from IPython.display import Image
Image(filename='../imgs/banner.png')

%load_ext pretty_jupyter

Authors:

• Bailey Passmore, Data Scientist, HRDAG
• Larry Barrett, Consultant, HRDAG

import pendulum

Monday, 17 March 2025 at 11:47 AM (PDT)

%%html

<style>
    #Styling {
        font-weight: bold;
        font-family: Helvetica;
    }
</style>

Goal¶

• What we have in dev-pre-restructure is working fine but we're on a short deadline and need to streamline the data. Let's make a barebones table of the core data we need for analysis, including:
  • event_no
  • date_occurred (w/ year_occurred)
  • event_type (a field derived from the initial event type in the data, either 911 call reporting gunfire or Shotspotter GDT alert)
  • event_location (service address if specified)
  • date_dispatched (is this the same as date of arrival?)
  • area (the result of mapping the reported police district to the police area)

Research questions we're working towards¶

1. Is dispatch reported at the same rate for all districts?
2. RE: Soundthinking / Brookings Institute claim that some 80% of gunfire events do not get reported by citizens - is that True in Chicago?
   • When SST alerts aren't matched to 911 calls, what is the typical disposition of such an alert?
   • When 911 calls aren't matched to SST alert, what is the typical disposition?

Time period covered¶

• Earliest date occurred included: '2021-01-01'
• Last date occurred included: '2024-09-23' (Per CPD, "ShotSpotter technology for the City of Chicago was discontinued in September 2024, meaning no records are available beyond that date.") The last date a ShotSpotter alert appears in the data is 23 September 2024, so we include records up to and including this date.

Identifying Event type (based on the _inittype field)¶

shotspotter = ['SST', 'PSST', 'MSST'] # keywords provided by CPD in Info sheet
citizencalls = ['SHOTS', 'SHOTSF', 'PERSHO',] # Note: 'PERGUN', 'PERDOW','PERHLP', 'DOMBAT', etc. excluded

Note about what humans can do that SST can't¶

In the OEMC data, in addition to reports originating as calls about shots fired ('SHOTSF') or persons being shot ('PERSHO'), we also see events with an initial label 'PERGUN' and final label referring to shots fired or someone shot. This tells us that not only do Chicagoans report gunfire in general, they may also report early warning signs of conflict involving firearms before any shots occur, giving first responders a head start to arrive on scene and provide potentially life-saving intervention.

# dependencies
import re
import numpy as np
from datetime import timedelta
import pandas as pd
from scipy.stats import power_divergence

# support methods
def format_count(v):
    return "{:,}".format(v)


def format_longfloat(prop):
    return float(f'{prop:.3f}')


def format_prop(prop, decn=1, asperc=True):
    if asperc: prop = prop*100
    return "{}%".format(round(prop, decn))


def report_fields(df, idcol, cols, fillna=False, headn=10):
    data = df[[idcol] + cols].drop_duplicates()
    if fillna:
        count = data[cols].fillna('None reported').value_counts().to_frame().reset_index()
        perc = data[cols].fillna('None reported').value_counts(normalize=True).to_frame().reset_index(
            ).rename(columns={'proportion': 'percent'})
    else:
        count = data[cols].value_counts().to_frame().reset_index()
        perc = data[cols].value_counts(normalize=True).to_frame().reset_index(
            ).rename(columns={'proportion': 'percent'})
    count['count'] = count['count'].apply(format_count)
    perc.percent = perc.percent.apply(format_prop)
    out = pd.merge(count, perc, on=cols)
    return out.head(headn)


def report_overtime(df, idcol, yearcol):
    yearly = df[[idcol, yearcol]].groupby(yearcol).nunique().reset_index()
    yearly[yearcol] = yearly[yearcol].astype(str)
    yearly = yearly.sort_values(yearcol)
    return yearly


def checkgroup(df):
    reporttrue = int(df.loc[df.dispatch_reported, 'count'].values[0].replace(',', ''))
    reportfalse = int(df.loc[~df.dispatch_reported, 'count'].values[0].replace(',', ''))
    return reportfalse > reporttrue


def lookup_cramer_cat(msrmt, df):
    if df == 1:
        if msrmt <= 0.1: return 'small'
        elif msrmt <= 0.3: return 'medium'
        else: return 'large'
    elif df == 2:
        if msrmt <= 0.07: return 'small'
        elif msrmt <= 0.21: return 'medium'
        else: return 'large'
    elif df == 3:
        if msrmt <= 0.06: return 'small'
        elif msrmt <= 0.17: return 'medium'
        else: return 'large'
    elif df == 4:
        if msrmt <= 0.05: return 'small'
        elif msrmt <= 0.15: return 'medium'
        else: return 'large'
    elif df == 5:
        if msrmt <= 0.04: return 'small'
        elif msrmt <= 0.13: return 'medium'
        else: return 'large'


def cramers(stat, n, df):
    msrmt = np.sqrt(stat/(n*df))
    cat = lookup_cramer_cat(msrmt=msrmt, df=df)
    return format_longfloat(msrmt), cat


def text_statistic(powerdiv, sig, n, degf, null_phrase, exp=None):
    rounded_stat = format_longfloat(powerdiv.statistic)
    rounded_pval = format_longfloat(powerdiv.pvalue)
    if rounded_pval == 0: rounded_pval = "< 0.0001"
    if exp: newinfo = f"This test results in a p-value of {rounded_pval}, "
    else: newinfo = f"This test results in a p-value of {rounded_pval}, "
    if powerdiv.pvalue < sig:
        effect_size, effect_cat = cramers(powerdiv.statistic, n, degf)
        newinfo += f"which is a statistically significant difference and rejects the null hypothesis that {null_phrase}. The test indicates a {effect_cat} effect. "
    else: newinfo += f"which is not statistically significant and fails to reject the null hypothesis that {null_phrase}. "
    if n < 30: newinfo += f"However, the sample size of {n} is small."
    return newinfo

# main
colorder = [
    'event_no',
    'date_occurred',
    'date_dispatched',
    'area',
    'location',
    #'location_x',
    #'location_y',
    'init_type',
    'event_type_init',
    'fin_type',
    'event_type_fin',
    'disposition',
    'event_type',
    'shotspotter_alert',
    'human_caller',
    'year_occurred',
    'dispatch_reported',
    'labeled_duplicate',
    'shotspotter_first',
    'disposition_reported',
    'misc_event',
    'early_warning',
    'any_coref',
    'cluster',
    'n_events',
]
data = pd.read_parquet("../../indicate/output/events.parquet")[colorder]
assert data.event_no.nunique() == data.shape[0]

sst = data.loc[data.event_type == 'ShotSpotter alert']
calls = data.loc[data.event_type == 'Human reporting gunfire']

Review data¶

Counts¶

Overall¶

• There are 385,448 gunfire-related 911 calls and ShotSpotter alerts prepared for this analysis.
• The data cover a time period between 2021-01-01 and 2024-09-22.

Source types¶

• Frequency table:

• Summary: Of the 385,448 emergency events included in the analysis,
  • 228,552 or 59.3% were generated by a 911 call, and
  • 156,896 or 40.7% were generated by a ShotSpotter alert.

Initial Event types¶

Presented are:

• the initial event type as reported by OEMC and CPD (init_type),
• the description of the initial type as found in the data (event_type_init), and
• the type of source which reported the event (event_type).

source_yearly = data[['event_no', 'year_occurred', 'event_type']
    ].groupby(['year_occurred', 'event_type']).nunique().reset_index()
src_ylr_piv = pd.pivot_table(
    source_yearly,
    values="event_no",
    index="year_occurred",
    columns="event_type",
    aggfunc="mean"
)
src_ylr_piv.plot(
    kind='bar',
    title='Emergency Events Observed Over Time',
    xlabel='Year Occurred', ylabel='Record count')

<Axes: title={'center': 'Emergency Events Observed Over Time'}, xlabel='Year Occurred', ylabel='Record count'>

Police Areas¶

Image(filename="../imgs/2019-CPD-Area-Boundaries.webp")

<IPython.core.display.Image object>

Referring to reporting and CPD public data, we map each police district observed in the data to the corresponding police area.

• Area 1, now called Area Central, will include the 2nd, 3rd, 7th, 8th, and 9th districts on the South Side.
• Area 2, now called Area South, will include the 4th, 5th, 6th, and 22nd districts on the Far South Side.
• Area 3, now called Area North, will include the 1st, 12th, 18th, 19th, 20th, and 24th districts on the North Side, largely along the lakefront.
• Area 4 will include the 10th, 11th, and 15th districts on the West Side.
• Area 5 will include the 14th, 16th, 17th, and 25th districts on the Northwest Side.

Events by police area¶

• The data refer to the police district associated with the call, but we map these to the reported corresponding police area to simplify analyses.
• Presented are the emergency event counts by police area.

Source type by area¶

• Presented are the record counts by source type and police area.

Area 1¶

• Presented are the record counts for Area 1, the area with the plurality of events, by source type and year occurred.

Research questions¶

%%jmd

Is dispatch reported at the same rate for all police areas?¶

Dispatch Date reporting¶

• Presented are the counts of observed non-missing "Dispatch Date" values.

disp_byarea = report_fields(
    df=data,
    idcol='event_no',
    cols=['dispatch_reported', 'area'], headn=10).reset_index(drop=True).sort_values([
    'area', 'dispatch_reported',]).reset_index(drop=True)
more_unreported = disp_byarea.groupby('area')[['dispatch_reported', 'count']
    ].apply(checkgroup).reset_index().rename(columns={0: 'more_unreported'})
more_unreported = pd.merge(disp_byarea, more_unreported, on='area')

Dispatch Date reporting by area¶

• Presented are the counts of observed non-missing "Dispatch Date" values by police area.

• Highlighting areas where the rate of missing "Dispatch Date" values is higher than the rate of non-missing values.

obsrate_byarea = data[['event_no', 'area',]].groupby('area')[[
    'event_no',]].apply(
    lambda df: df.event_no.nunique()/data.shape[0]
    ).reset_index().rename(columns={0: 'prop_events_reported'})
disprate_byarea = data[['event_no', 'area', 'dispatch_reported']
    ].groupby('area')[[
    'event_no', 'dispatch_reported']].apply(
    lambda df: df.dispatch_reported.sum()/df.event_no.nunique()
    ).reset_index().rename(columns={0: 'prop_dispatch_reported'})

# Assuming equal likelihood
cats = disprate_byarea.area.unique()
total = data.area.notna().sum()
props = disprate_byarea.prop_dispatch_reported
sig=0.05
loglikely = power_divergence(props, lambda_='log-likelihood')
chisquare = power_divergence(props, lambda_='pearson')

base_info = f"""This test will compare whether observed racial proportions match the proportion of each racial group in \
the general population. In interpreting the results, a p-value below {sig} will be considered statistically significant. \
In such cases where the p-value is statistically significant, the effect size will be measured and categorized as small, medium, or large.\n"""
null_phrase="the distribution of _____ follows the distribution of _____"
loglikely_info = text_statistic(
    loglikely,
    sig=sig, n=total, degf=len(cats)-1-0,
    null_phrase=null_phrase)
chisquare_info = text_statistic(
    chisquare,
    sig=sig, n=total, degf=len(cats)-1-0,
    null_phrase=null_phrase)

print(base_info)
print(chisquare_info)

This test will compare whether observed racial proportions match the proportion of each racial group in the general population. In interpreting the results, a p-value below 0.05 will be considered statistically significant. In such cases where the p-value is statistically significant, the effect size will be measured and categorized as small, medium, or large.

This test results in a p-value of 1.0, which is not statistically significant and fails to reject the null hypothesis that the distribution of _____ follows the distribution of _____. 

Summary¶

• The Chi-Square test results tell us that the police areas have a similar proportion of dispatch reported when compared to each other.
• Let's dig deeper and see if that is still true regardless of the type of source that reported the event (ShotSpotter alert or human caller).

Dispatch Date reporting by area and source type¶

• Presented are the counts of observed non-missing "Dispatch Date" values by police area and the type of source which reported the emergency.

Dispatch Date reporting by area and source type¶

• Presented are the counts of observed non-missing "Dispatch Date" values by police area, source type, and

Area 1¶

Area 2¶

Area 3¶

Area 4¶

Area 5¶

Dispatch Date reporting by area and source type¶

• Presented are the counts of observed non-missing "Dispatch Date" values by police area, source type, and

ShotSpotter alerts¶

911 calls¶

Summary¶

RE: Soundthinking / Brookings Institute claim that some 80% of gunfire events do not get reported by citizens - is that True in Chicago?¶

notmisc = sst.loc[(sst.misc_event == False) & (sst.disposition_reported)]

Disposition data¶

In the info page included with the data, the CPD FOIA officer informed us that they had internally identified emergency events from both sources, 911 callers and ShotSpotter, that referred to the same underlying gunfire event, and that the disposition field was only included in the responsive records when ShotSpotter was the first to report. Source

• Of the 385,448 emergency events included in the analysis, 123,169 or 78.5% have a reported disposition.

ShotSpotter data, labeled duplicate

ShotSpotter data, disposition provided

ShotSpotter data, labeled duplicate and disposition provided

5 most frequently reported disposition values¶

• Of the 123,169 emergency events about potential gunfire identified by CPD as first reported by ShotSpotter, 104,550 or 84.9% are labeled as a "Miscellaneous Incident."
• 84.9% of ShotSpotter-original alerts are labeled as a "Miscellaneous Incident".

5 most frequently reported disposition values other than "Miscellaneous Incident"¶

Statements and assumptions¶

Underlying events

• The real-world event involving gunfire that prompted a true positive report to OEMC and/or CPD will be referred to as the "underlying emergency event". These emergency records have not been de-duplicated with respect to the singular real-world event being reported, and it's possible that multiple sources provided a report for any given event.
• In producing a single ShotSpotter alert, Soundthinking has described their service as triangulating detections from multiple sensors/microphones to pinpoint the location and produce an alert to send to law enforcement and emergency services. To that point, we might expect that no two ShotSpotter alerts may be co-referant for a singular event, ie) ShotSpotter does not send additional alerts if law enforcement has not yet arrived on scene. However, the same would not be true for 911 callers, for whom there is reporting describing repeat calls to 911 after time passed without any officers arriving on scene to address the emergency.

Linking reports of the same event

• Recall that per CPD, disposition was only reported when ShotSpotter was the first to alert OEMC and/or CPD to the emergency. If ShotSpotter was not the first to alert OEMC and/or CPD to an underlying emergency, then I assume that the first alert came from a human (in other words and most often, from a civilian 911 call). Additionally, regardless of whether or not the alert came first, CPD labeled all ShotSpotter alerts indicating whether or not the alert was understood to be a duplicate for another report.
  • If ShotSpotter was not the first to alert OEMC and/or CPD to a potential gunfire event, then the ShotSpotter alert is co-referant to at least one 911 call for the same underlying emergency event.
• Two emergency event records will be considered "co-referant" if:
  1. the date_occurred timestamps are within +/- 10 minutes of each other, and
  2. the location values are at least 95% similar in terms of string distance. Note that this method has limitations and will likely only identify events on streets of the same name as co-referant, rather than linking events that are geographically neighbors of each other.

Interpreting record linkage by case

• Case 1: ShotSpotter alerts with no co-referent 911 calls
  • labeled_duplicate should be False
  • We should have disposition reported for all of these events in the data.
  • When disposition refers to a "Miscellaneous Incident", we will consider the alert to be a false positive gunfire alert, and all other disposition categories will be assumed true positive gunfire alert.
  • The proportion of alerts that have no co-referant 911 calls and a disposition other than "Miscellaneous Incident" will be interpreted as the proportion of gunfire events captured by ShotSpotter that would have otherwise gone unreported to OEMC and/or CPD.
    • This proportion will be compared to 1) the Brookings Institute proposed statistic of 80% of gunfire events and 2) the proportion of false positive ShotSpotter gunfire alerts.
• Case 2: ShotSpotter alerts with co-referent 911 calls
  • labeled_duplicate may or may not be False
  • We should have disposition reported for those events which ShotSpotter alertted to first, but only those.
  • We have no disposition data to assess the proportion of 911 calls that are false positive gunfire reports. However, if a ShotSpotter alert is co-referant with a 911 call, then either 1) both reports are true positives or 2) both are false positives, so the proportion is irrelevant.
• Case 3: 911 calls with no co-referent ShotSpotter alerts
  • labeled_duplicate will not be reported for the vast majority of these events
  • No disposition data reported for these events to assess the proportion of 911 calls that are false positive gunfire reports, but the true positives would be interpreted as the proportion of gunfire events not captured by ShotSpotter that would have otherwise gone unreported to OEMC and/or CPD.

Tests¶

If the Brookings Institute's proposed statistic is reasonable to apply in Chicago, we expect that no more than 20% of underlying gunfire reports in the data are one of the following:

• Subset of Case 2: a 911 call which occurs prior to a co-referent ShotSpotter alert
• Case 3: a 911 call with no co-referent ShotSpotter alert

Alternatively, we expect that the proportion of ShotSpotter alerts that have no co-referant 911 calls and a disposition other than "Miscellaneous Incident" will be some 80% of underlying gunfire reports.

sst_notdup = sst.loc[(sst.labeled_duplicate == False)]

sst_nocands = (sst.cluster.isna())
sst_nocands_vc = report_fields(
    df=sst,
    idcol='event_no',
    cols=['shotspotter_alert', 'any_coref', 'misc_event',],
    headn=5)

• The data we obtained from CPD only covered ShotSpotter alerts, so we only have labels from CPD indicating duplicate reports for those events that Shotspotter at least one alert.

Emergency events by reporting source and whether CPD labeled the event as a duplicate report¶

ShotSpotter alerts labeled as NOT a duplicate event, by multiple factors¶

• 17,150 or 13.9% of the emergency events analyzed were ShotSpotter alerts that have no identified co-referant 911 calls and a disposition referring to something other than "Miscellaneous Incident."

sst[['shotspotter_first', 'disposition_reported', 'misc_event']].fillna('no data').value_counts()

shotspotter_first  disposition_reported  misc_event
True               True                  True          104550
False              False                 no data        33727
True               True                  False          18619
Name: count, dtype: int64

import plotly.graph_objects as go

fig = go.Figure(data=[go.Sankey(
    node = dict(
      pad = 15,
      thickness = 20,
      line = dict(color = "black", width = 0.5),
      label = ["Shotspotter alerts",
               "Labeled first report",
               "Not labeled first report (no disposition data)",
               "Labeled miscellaneous incident",
               "Not labeled miscellaneous incident",
              ],
      color = ["#00cc7a", "#00ccad", "green", "#00b8cc", "#00cc7a",]
    ),
    link = dict(
      source = [0, 0, 1, 2, 1, 2,], # indices correspond to labels, eg A1, A2, A1, B1, ...
      target = [1, 2, 3, 3, 4, 4,],
      value = [
          (sst.shotspotter_first == True).sum(),
          (sst.shotspotter_first == False).sum(),
          ((sst.shotspotter_first == True) & (sst.misc_event == True)).sum(),
          ((sst.shotspotter_first == False) & (sst.misc_event == True)).sum(),
          ((sst.shotspotter_first == True) & (sst.misc_event == False)).sum(),
          ((sst.shotspotter_first == False) & (sst.misc_event == False)).sum(),
      ],
      hovercolor=["#00cc7a", "#00ccad", "lightblue", "mediumturquoise", "lightgreen", "cyan",],
  ))])

fig.update_layout(title_text="Labeling of ShotSpotter alerts", font_size=10)
fig.show()

Summary¶